Import the Libraries¶
In [1]:
import os
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import plotly.graph_objs as go
import plotly.subplots as sp
import numpy as np
from scipy.optimize import minimize
from DP_Binary import *
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from DQN_Binary.PortfolioRebalancerEnv_Roy import PortfolioRebalancerEnv
import plotly.graph_objs as go
import plotly.subplots as sp
import plotly.express as px
from sklearn.metrics import confusion_matrix
sns.set_theme()
In [2]:
plt.rcParams.update({"font.size": 18})
tmp_device = torch.device("cpu")
In [3]:
og_path = os.getcwd()
og_path
Out[3]:
'/home/guest/Roy'
Initialize¶
In [4]:
n_asset = 2
In [5]:
mu = np.linspace(50, 200, n_asset) / 1e4
sigma = np.linspace(300, 800, n_asset) / 1e4
cov = np.diag(sigma ** 2)
In [6]:
tc_list = [0, 0.0005, 0.001, 0.002, 0.005, 0.01, 0.05]
In [7]:
x = np.arange(1, 101)
state_possible = np.array(np.meshgrid(*([x] * n_asset))).T.reshape(-1, n_asset).astype(np.float32)
state_possible = state_possible[state_possible.sum(axis=1) == 100, :] / 100
In [8]:
def net_sharpe(w1, mu, cov, w0, tc):
return (w1.dot(mu) - cost_turnover(w0, w1, tc)) / np.sqrt(w1.dot(cov).dot(w1.T))
def obj_func(x, mu, cov):
return -x.dot(mu) / np.sqrt(x.dot(cov).dot(x))
def find_optimal_wgt(mu, cov):
n = len(mu)
w_min = np.zeros(n)
w_max = np.ones(n) * 2 / n
x0 = np.ones(n) / n
bounds = np.vstack([w_min, w_max]).T
cstr = [{"type": "eq", "fun": lambda x: np.sum(x) - 1, "jac": lambda x: np.ones(n)}]
opt = minimize(fun=obj_func, x0=x0, args=(mu, cov),
bounds=bounds,
constraints=cstr,
tol=1e-6,
options={"maxiter": 10000})
if not opt.success:
raise ValueError("optimization failed: {}".format(opt.message))
return opt.x / opt.x.sum()
def cost_turnover(w0, w1, tc):
return np.sum(np.abs(w1 - w0) * tc) / 2
def expected_cost_total(w0, w1, opt_w, mu, cov, tc):
opt_net_sharpe = net_sharpe(opt_w, mu, cov, w0, tc)
w1_net_sharpe = net_sharpe(w1, mu, cov, w0, tc)
return opt_net_sharpe - w1_net_sharpe
DP¶
In [9]:
dp_path = og_path + '/DP/results/models/'
os.chdir(dp_path)
In [10]:
df_0 = pd.read_csv("dp_tc_0_assets_2", index_col=0, dtype=np.float32)
df_01 = pd.read_csv("dp_tc_0.001_assets_2", index_col=0, dtype=np.float32)
df_02 = pd.read_csv("dp_tc_0.002_assets_2", index_col=0, dtype=np.float32)
df_05 = pd.read_csv("dp_tc_0.0005_assets_2", index_col=0, dtype=np.float32)
df_5 = pd.read_csv("dp_tc_0.005_assets_2", index_col=0, dtype=np.float32)
df_1 = pd.read_csv("dp_tc_0.01_assets_2", index_col=0, dtype=np.float32)
df_5 = pd.read_csv("dp_tc_0.05_assets_2", index_col=0, dtype=np.float32)
In [11]:
df_list = [df_0, df_05, df_01, df_02, df_5, df_1, df_5]
In [12]:
x = pd.to_numeric(df_0.index.values)
action_possible = pd.to_numeric(df_0.columns.values)
In [13]:
action_df = pd.DataFrame(index=x)
action_bm_df = pd.DataFrame(index=x)
for tc, bell in zip(tc_list, df_list):
# visualize q table
action = np.array([action_possible[i] for i in np.argmax(bell, axis=1)])
action_df[f"TC: {tc * 1e4:.0f} bps BM"] = action
action_bm = []
for i in x:
a = x[np.argmax([net_sharpe(np.array([j, 1 - j]), mu, cov, np.array([i, 1 - i]), tc) for j in x])] - i
action_bm.append(a)
action_bm = np.array(action_bm)
action_bm_df[f"TC: {tc * 1e4:.0f} bps BM (GT)"] = action_bm
In [14]:
optimal_weight = find_optimal_wgt(mu, cov)
optimal_weight
Out[14]:
array([0.63941273, 0.36058727])
In [15]:
fig, ax = plt.subplots(1, 1, figsize=(20, 10))
action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'])
action_bm_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dashed", alpha=0.5)
ax.set_xlabel("Weight on First Asset")
ax.set_ylabel("Suggested Delta Weight on Asset 1")
ax.axvline(optimal_weight[0], color="red", linestyle="dotted")
ax.axhline(0, color="red", linestyle="dotted")
ax.legend()
plt.tight_layout()
plt.show()
plt.close()
In [16]:
fig, ax = plt.subplots(1, 1, figsize=(20, 10))
action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'])
action_bm_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dashed", alpha=0.5)
ax.set_xlabel("Weight on First Asset")
ax.set_ylabel("Suggested Delta Weight on Asset 1")
ax.axvline(optimal_weight[0], color="red", linestyle="dotted")
ax.axhline(0, color="red", linestyle="dotted")
ax.legend()
plt.tight_layout()
plt.show()
plt.close()
In [17]:
# Create subplots
fig = sp.make_subplots(rows=1, cols=1, subplot_titles="Comparison of BM and Net Sharpe State Possible Outputs", shared_yaxes=True)
# Add traces for Action DQN and Action BM
for col, color, linestyle in zip(action_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['solid']*len(action_df.columns)):
trace_dqn = go.Scatter(x=action_df.index, y=action_df[col], mode='lines', name=f'Action BM - {col}', line=dict(color=color, dash="solid"))
fig.add_trace(trace_dqn, row=1, col=1)
# Add traces for Action DQN and Action BM
for col, color, linestyle in zip(action_bm_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['dot']*len(action_bm_df.columns)):
trace_bm = go.Scatter(x=action_bm_df.index, y=action_bm_df[col], mode='lines', name=f'Ground Truth - {col}', line=dict(color=color, dash=linestyle))
fig.add_trace(trace_bm, row=1, col=1)
# Add vertical and horizontal lines
fig.add_shape(dict(
type="line",
x0=optimal_weight[0],
x1=optimal_weight[0],
y0=action_df.values.min(),
y1=action_df.values.max(),
line=dict(color="gold", dash="dot")
))
fig.add_shape(dict(
type="line",
x0=action_df.index.min(),
x1=action_df.index.max(),
y0=0,
y1=0,
line=dict(color="gold", dash="dot")
))
# Update layout
fig.update_layout(
title="Comparison of BM and Net Sharpe State Possible Outputs",
xaxis=dict(title="Weight on First Asset"),
yaxis=dict(title="Suggested Delta Weight on Asset 1"),
showlegend=True,
)
# Show the figure
fig.show()
- Change DP ground truth to run on binary (check)
- DQN still weird need debug with higher TC (in progress)
- Decision boundary changed so need to re-run DP (check)
- Use Bar Graph for plotting (TBD)
DP Binary¶
In [18]:
dp_path = og_path + '/DP_Binary/results/models/'
os.chdir(dp_path)
In [19]:
df_0 = pd.read_csv("dp_tc_0_assets_2", index_col=0, dtype=np.float32)
df_01 = pd.read_csv("dp_tc_0.001_assets_2", index_col=0, dtype=np.float32)
df_02 = pd.read_csv("dp_tc_0.002_assets_2", index_col=0, dtype=np.float32)
df_05 = pd.read_csv("dp_tc_0.0005_assets_2", index_col=0, dtype=np.float32)
df_5 = pd.read_csv("dp_tc_0.005_assets_2", index_col=0, dtype=np.float32)
df_1 = pd.read_csv("dp_tc_0.01_assets_2", index_col=0, dtype=np.float32)
df_5 = pd.read_csv("dp_tc_0.05_assets_2", index_col=0, dtype=np.float32)
In [20]:
df_list = [df_0, df_05, df_01, df_02, df_5, df_1, df_5]
In [21]:
x = pd.to_numeric(df_0.index.values)
action_possible = pd.to_numeric(df_0.columns.values)
In [22]:
action_df = pd.DataFrame(index=x)
action_bm_df = pd.DataFrame(index=x)
for tc, bell in zip(tc_list, df_list):
# visualize q table
action = np.array([action_possible[i] for i in np.argmax(bell, axis=1)])
action_df[f"TC: {tc * 1e4:.0f} bps BM"] = action
action_bm = []
for i in x:
a = x[np.argmax([net_sharpe(np.array([j, 1 - j]), mu, cov, np.array([i, 1 - i]), tc) for j in x])] - i
action_bm.append(a)
action_bm = np.array(action_bm)
action_bm_df[f"TC: {tc * 1e4:.0f} bps BM (GT)"] = action_bm
In [23]:
optimal_weight = find_optimal_wgt(mu, cov)
optimal_weight
Out[23]:
array([0.63941273, 0.36058727])
In [24]:
fig, ax = plt.subplots(1, 1, figsize=(20, 10))
action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'])
action_bm_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dashed", alpha=0.5)
ax.set_xlabel("Weight on First Asset")
ax.set_ylabel("Suggested Delta Weight on Asset 1")
ax.axvline(optimal_weight[0], color="red", linestyle="dotted")
ax.axhline(0, color="red", linestyle="dotted")
ax.legend()
plt.tight_layout()
plt.show()
plt.close()
In [25]:
fig, ax = plt.subplots(1, 1, figsize=(20, 10))
action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'])
action_bm_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dashed", alpha=0.5)
ax.set_xlabel("Weight on First Asset")
ax.set_ylabel("Suggested Delta Weight on Asset 1")
ax.axvline(optimal_weight[0], color="red", linestyle="dotted")
ax.axhline(0, color="red", linestyle="dotted")
ax.legend()
plt.tight_layout()
plt.show()
plt.close()
In [26]:
# Create subplots
fig = sp.make_subplots(rows=1, cols=1, subplot_titles="Comparison of BM and Net Sharpe State Possible Outputs", shared_yaxes=True)
# Add traces for Action DQN and Action BM
for col, color, linestyle in zip(action_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['solid']*len(action_df.columns)):
trace_dqn = go.Scatter(x=action_df.index, y=action_df[col], mode='lines', name=f'Action BM - {col}', line=dict(color=color, dash="solid"))
fig.add_trace(trace_dqn, row=1, col=1)
# Add traces for Action DQN and Action BM
for col, color, linestyle in zip(action_bm_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['dot']*len(action_bm_df.columns)):
trace_bm = go.Scatter(x=action_bm_df.index, y=action_bm_df[col], mode='lines', name=f'Ground Truth - {col}', line=dict(color=color, dash=linestyle))
fig.add_trace(trace_bm, row=1, col=1)
# Add vertical and horizontal lines
fig.add_shape(dict(
type="line",
x0=optimal_weight[0],
x1=optimal_weight[0],
y0=action_df.values.min(),
y1=action_df.values.max(),
line=dict(color="gold", dash="dot")
))
fig.add_shape(dict(
type="line",
x0=action_df.index.min(),
x1=action_df.index.max(),
y0=0,
y1=0,
line=dict(color="gold", dash="dot")
))
# Update layout
fig.update_layout(
title="Comparison of BM and Net Sharpe State Possible Outputs",
xaxis=dict(title="Weight on First Asset"),
yaxis=dict(title="Suggested Delta Weight on Asset 1"),
showlegend=True,
)
# Show the figure
fig.show()
DDPG¶
TODO: Need to run DDPG for two asset case only¶
In [27]:
# ddpg_path = og_path + '/DDPG/'
# os.chdir(ddpg_path)
In [28]:
# from DDPG import *
In [29]:
# ddpg_action_df = pd.DataFrame(index=state_possible[:, 0])
# for tc in tc_list:
# agent = Agent(alpha=0.01, beta=0.01, input_dims=[n_asset], tau=0.01, tc=tc,
# batch_size=100000, layer1_size=400, layer2_size=300, n_actions=n_asset)
# agent.load_models()
# action = []
# for current_state in state_possible:
# tmp_action = agent.choose_action(current_state)
# action.append(tmp_action[0]) # just store action of first asset
# action = np.array(action)
# ddpg_action_df[f"TC: {tc * 1e4:.0f} bps"] = action
In [30]:
# ddpg_action_df
DQN Binary¶
In [31]:
class DQN(nn.Module):
def __init__(self, n_observations, n_actions, name, tc, chkpt_dir='results/models'):
super(DQN, self).__init__()
self.layer1 = nn.Linear(n_observations, 32)
self.layer2 = nn.Linear(32, 32)
self.layer3 = nn.Linear(32, n_actions)
self.checkpoint_file = os.path.join(chkpt_dir, name + f'_tc_{tc}_assets_{n_observations}_dqn')
# Called with either one element to determine next action, or a batch
# during optimization. Returns tensor([[left0exp,right0exp]...]).
def forward(self, x):
x = F.relu(self.layer1(x))
x = F.relu(self.layer2(x))
return self.layer3(x)
def save_checkpoint(self):
print('... saving checkpoint ...')
torch.save(self.state_dict(), self.checkpoint_file)
def load_checkpoint(self):
print('... loading checkpoint ...')
self.load_state_dict(torch.load(self.checkpoint_file))
In [32]:
def save_models(pn, tn):
pn.save_checkpoint()
tn.save_checkpoint()
def load_models(pn, tn):
pn.load_checkpoint()
tn.load_checkpoint()
In [33]:
ddpg_path = og_path + '/DQN_Binary/'
os.chdir(ddpg_path)
In [34]:
optimal_weight = find_optimal_wgt(mu, cov)
optimal_weight
Out[34]:
array([0.63941273, 0.36058727])
In [35]:
env = PortfolioRebalancerEnv(
mu=mu,
sigma=cov,
w_optimal=optimal_weight,
n_assets=n_asset,
)
# device = torch.device("cpu")
In [36]:
BATCH_SIZE = 128
GAMMA = 0.9
EPS_START = 0.9999
EPS_END = 0.05
EPS_DECAY = 1_000
TAU = 0.005
LR = 1e-4
mem_cap = 10_000
mode = 'FF'
num_episodes = 50_000
In [37]:
x = np.arange(1, 101)
state_possible = np.array(np.meshgrid(*([x] * 2))).T.reshape(-1, 2).astype(
np.float32)
state_possible = state_possible[state_possible.sum(axis=1) == 100, :] / 100
In [38]:
dqn_bool_action_df = pd.DataFrame(index=x)
In [39]:
x = state_possible[:, 0]
dqn_action_df = pd.DataFrame(index=x)
for tc in tc_list:
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
# visualize q table
action = []
for j in range(x.shape[0]):
qval = a(torch.FloatTensor(state_possible[j])).detach().numpy()
action.append(env.action_space[qval.argmax()])
action = np.array(action)
dqn_action_df[f"TC: {tc * 1e4:.0f} bps"] = action
dqn_action_df
... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ...
Out[39]:
| TC: 0 bps | TC: 5 bps | TC: 10 bps | TC: 20 bps | TC: 50 bps | TC: 100 bps | TC: 500 bps | |
|---|---|---|---|---|---|---|---|
| 0.01 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| 0.02 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| 0.03 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| 0.04 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| 0.05 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| ... | ... | ... | ... | ... | ... | ... | ... |
| 0.95 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.96 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.97 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.98 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.99 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
99 rows × 7 columns
In [40]:
fig, ax = plt.subplots(1, 1, figsize=(20, 10))
action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dashed")
dqn_action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dotted")
ax.set_xlabel("Weight on First Asset")
ax.set_ylabel("Suggested Delta Weight on Asset 1")
ax.axvline(optimal_weight[0], color="red", linestyle="dotted")
ax.axhline(0, color="red", linestyle="dotted")
ax.legend()
plt.tight_layout()
plt.show()
Visualize Results¶
In [41]:
sns.heatmap(dqn_action_df, cmap='cividis')
y = optimal_weight[0]
ax = plt.gca()
plt.hlines(y=y*100, xmin=0, xmax=len(dqn_action_df.columns), colors='red')
plt.title("DQN Binary Heatmap of Rebalancing")
plt.show()
In [42]:
# Assuming `dqn_bool_action_df`, `action_df`, and `optimal_weight` are available
# Create subplots
fig = sp.make_subplots(rows=1, cols=1, subplot_titles=("Comparison of DQN and Action"), shared_yaxes=True)
# Add traces for DQN and Action
for col, color, linestyle in zip(dqn_action_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['solid']*len(dqn_action_df.columns)):
trace_dqn = go.Scatter(x=dqn_action_df.index, y=dqn_action_df[col], mode='lines', name=f'DQN - {col}', line=dict(color=color, dash=linestyle))
fig.add_trace(trace_dqn, row=1, col=1)
# Add traces for DQN and Action
for col, color, linestyle in zip(action_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['dash']*len(action_df.columns)):
trace_action = go.Scatter(x=action_df.index, y=action_df[col], mode='lines', name=f'BM - {col}', line=dict(color=color, dash=linestyle))
fig.add_trace(trace_action, row=1, col=1)
# Add traces for Action DQN and Action BM
for col, color, linestyle in zip(action_bm_df.columns, ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#2a8bde', '#ff6e51', '#4caf50'], ['dot']*len(action_bm_df.columns)):
trace_bm = go.Scatter(x=action_bm_df.index, y=action_bm_df[col], mode='lines', name=f'Ground Truth - {col}', line=dict(color=color, dash=linestyle))
fig.add_trace(trace_bm, row=1, col=1)
# Add vertical and horizontal lines
fig.add_shape(dict(
type="line",
x0=optimal_weight[0],
x1=optimal_weight[0],
y0=action_df.values.min(),
y1=action_df.values.max(),
line=dict(color="gold", dash="dot")
))
fig.add_shape(dict(
type="line",
x0=dqn_action_df.index.min(),
x1=dqn_action_df.index.max(),
y0=0,
y1=0,
line=dict(color="gold", dash="dot")
))
# Update layout
fig.update_layout(
title="Comparison of DQN and Action",
xaxis=dict(title="Weight on First Asset"),
yaxis=dict(title="Suggested Delta Weight on Asset 1"),
showlegend=True,
)
# Show the figure
fig.show()
In [43]:
# Assuming `dqn_bool_action_df` and `bool_action_bm_df` are your DataFrames
dqn_df = dqn_action_df.copy()
bm_df = action_df.copy()
# Create a heatmap trace for DQN Binary
trace_dqn = go.Heatmap(z=dqn_df, x=dqn_df.columns, y=dqn_df.index, name="DQN Binary")
# Create a heatmap trace for DP Bellman
trace_dp = go.Heatmap(z=bm_df, x=bm_df.columns, y=bm_df.index, name="DP Bellman")
# Create a heatmap trace for Q Table
# trace_q = go.Heatmap(z=q_df, x=q_df.columns, y=q_df.index, name="Q Table")
# Create the layout with 'overlay' parameter set to "overlay"
layout = go.Layout(
title="Comparison of DQN Binary and DP Bellman",
xaxis=dict(title="Transaction Cost (bps)",
showgrid=True),
yaxis=dict(title="Weight on First Asset",
showgrid=True))
# Add horizontal line at y=0.50
layout.shapes = [
dict(
type="line",
x0=dqn_df.columns[0], # Adjust x0 to the starting point of your data
x1=dqn_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
),
dict(
type="line",
x0=bm_df.columns[0], # Adjust x0 to the starting point of your data
x1=bm_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
),
# dict(
# type="line",
# x0=q_df.columns[0], # Adjust x0 to the starting point of your data
# x1=q_df.columns[-1], # Adjust x1 to the ending point of your data
# y0=optimal_weight[0],
# y1=optimal_weight[0],
# line=dict(color="red", width=2))
]
# Create the figure
fig = go.Figure(data=[trace_dp], layout=layout)
fig.show()
fig = go.Figure(data=[trace_dqn], layout=layout)
fig.show()
# fig = go.Figure(data=[trace_q], layout=layout)
# fig.show()
In [44]:
# Create an empty dataframe to store the confusion matrices
confusion_matrices = {}
# Iterate through the columns and calculate confusion matrices
for col1, col2 in zip(bm_df.columns, dqn_df.columns):
true_labels = bm_df[col1]
predicted_labels = dqn_df[col2]
# Create the confusion matrix
confusion = confusion_matrix(true_labels, predicted_labels)
confusion_matrices[col1] = confusion
# Print or use the confusion matrices as needed
for column, confusion in confusion_matrices.items():
print(f"Confusion Matrix for {column}:")
print(confusion)
Confusion Matrix for TC: 0 bps BM: [[ 1 0] [ 2 96]] Confusion Matrix for TC: 5 bps BM: [[ 2 3] [ 3 91]] Confusion Matrix for TC: 10 bps BM: [[ 8 1] [ 0 90]] Confusion Matrix for TC: 20 bps BM: [[21 0] [ 1 77]] Confusion Matrix for TC: 50 bps BM: [[81 18] [ 0 0]] Confusion Matrix for TC: 100 bps BM: [[90 0] [ 9 0]] Confusion Matrix for TC: 500 bps BM: [[99]]
In [45]:
dqn_noreb = {}
bm_noreb = {}
for col1, col2 in zip(bm_df.columns, dqn_df.columns):
bm_index = bm_df[bm_df[col1] == 0].index.values
dqn_index = dqn_df[dqn_df[col2] == 0].index.values
bm_noreb[col2] = bm_index
dqn_noreb[col2] = dqn_index
In [46]:
# Create a DataFrame from the dictionary, excluding keys with empty values
df = pd.DataFrame.from_dict(dqn_noreb, orient='index').T
# Create a box plot using Plotly
fig = px.box(df, title="Box Plot of Data")
# Show the plot
fig.show()
In [47]:
# Create a DataFrame from the dictionary, excluding keys with empty values
df = pd.DataFrame.from_dict(bm_noreb, orient='index').T
# Create a box plot using Plotly
fig = px.box(df, title="Box Plot of Data")
# Show the plot
fig.show()
Higher Dimensions¶
4 Assets¶
In [48]:
ddpg_path = og_path + '/DQN_Binary/'
os.chdir(ddpg_path)
In [49]:
n_asset = 4
tc = 0.002
mu = np.linspace(50, 200, n_asset) / 1e4
sigma = np.linspace(300, 800, n_asset) / 1e4
cov = np.diag(sigma ** 2)
In [50]:
optimal_weight = find_optimal_wgt(mu, cov)
state = torch.Tensor(optimal_weight).unsqueeze(0)
optimal_weight
Out[50]:
array([0.32646603, 0.26983215, 0.21982399, 0.18387782])
In [51]:
x = np.arange(1, 101)
spn = np.array(np.meshgrid(*([x] * len(mu)))).T.reshape(-1, len(mu)).astype(
np.float32)
spn = spn[spn.sum(axis=1) == 100, :] / 100
tc_list = [0, 0.0005, 0.001, 0.002, 0.005, 0.01, 0.05]
In [52]:
x = state_possible[:, 0]
dqn_action_df = pd.DataFrame(index=x)
for tc in tc_list:
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
# visualize q table
action = []
for j in range(x.shape[0]):
qval = a(torch.FloatTensor(spn[j])).detach().numpy()
action.append(env.action_space[qval.argmax()])
action = np.array(action)
dqn_action_df[f"TC: {tc * 1e4:.0f} bps"] = action
dqn_action_df
... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ...
Out[52]:
| TC: 0 bps | TC: 5 bps | TC: 10 bps | TC: 20 bps | TC: 50 bps | TC: 100 bps | TC: 500 bps | |
|---|---|---|---|---|---|---|---|
| 0.01 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.02 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.03 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.04 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.05 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| ... | ... | ... | ... | ... | ... | ... | ... |
| 0.95 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.96 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.97 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.98 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.99 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
99 rows × 7 columns
In [53]:
fig, ax = plt.subplots(1, 1, figsize=(20, 10))
# ql_action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'])
# bell_action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dashed")
dqn_action_df.plot(ax=ax, color=['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728'], linestyle="dotted")
ax.set_xlabel("Weight on First Asset")
ax.set_ylabel("Suggested Delta Weight on Asset 1")
ax.axvline(optimal_weight[0], color="red", linestyle="dotted")
ax.axhline(0, color="red", linestyle="dotted")
ax.legend()
plt.tight_layout()
plt.show()
In [54]:
def fix_reward(x):
return [i.cpu().item() for i in x]
In [55]:
# reward_df = pd.DataFrame()
# for col, val in reward_dict.items():
# reward_df[col] = fix_reward(reward_dict[col])
# reward_df
In [56]:
sns.heatmap(dqn_action_df, cmap='cividis')
y = optimal_weight[0]
ax = plt.gca()
plt.hlines(y=y*100, xmin=0, xmax=len(dqn_action_df.columns), colors='red')
plt.title("DQN Binary Heatmap of Rebalancing")
plt.show()
In [57]:
# Assuming `dqn_bool_action_df` and `bool_action_bm_df` are your DataFrames
dqn_df = dqn_action_df.copy()
# Create a heatmap trace for DQN Binary
trace_dqn = go.Heatmap(z=dqn_df, x=dqn_df.columns, y=dqn_df.index, name="DQN Binary")
# Create the layout with 'overlay' parameter set to "overlay"
layout = go.Layout(
title=f"DQN Binary for {n_asset} case",
xaxis=dict(title="Transaction Cost (bps)",
showgrid=True),
yaxis=dict(title="Weight on First Asset",
showgrid=True))
# Add horizontal line at y=0.50
layout.shapes = [
dict(
type="line",
x0=dqn_df.columns[0], # Adjust x0 to the starting point of your data
x1=dqn_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
)
]
fig = go.Figure(data=[trace_dqn], layout=layout)
fig.show()
In [58]:
px.line(optimal_weight).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [59]:
tc = 0.002
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
... loading checkpoint ... ... loading checkpoint ...
In [60]:
qval = a(torch.FloatTensor(optimal_weight)).detach().numpy()
act = env.action_space[qval.argmax()]
if act.item() == 1:
print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
else:
print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
Action = 0 -> You should NOT rebalance! (tc = 20 bps)
In [61]:
import torch
def get_high_dim_tensors(og_tensor, w_optimal, num_steps=10):
if type(og_tensor) != torch.Tensor:
og_tensor = torch.Tensor(og_tensor).unsqueeze(0)
# Calculate the increase and decrease values
num_steps_up = num_steps // 2
num_steps_down = num_steps - num_steps_up
# Calculate the increase and decrease values
increase_values = torch.linspace(0, 1 - og_tensor[0, 0], num_steps_up)
decrease_values = torch.linspace(1 - og_tensor[0, 0], 0, num_steps_down)
# Create a list of tensors
result_list = []
a = torch.zeros(og_tensor.size(1))
a[0] = 1
b = torch.full((og_tensor.size(1),), 1 / (og_tensor.size(1) - 1))
b[0] = 0
c = torch.full((og_tensor.size(1),), 1 / (og_tensor.size(1)))
result_list.append(torch.Tensor(w_optimal))
result_list.append(a)
result_list.append(b)
result_list.append(c)
# above given state
for i in range(num_steps_up):
updated_tensor = og_tensor.clone()
updated_tensor[0, 0] += increase_values[i]
updated_tensor[0, 1:] -= (increase_values[i] / (og_tensor.size(1) - 1))
if torch.any(torch.lt(updated_tensor, 0)):
break
elif torch.equal(updated_tensor.squeeze(), torch.Tensor(w_optimal)):
continue
else:
result_list.append(updated_tensor.squeeze())
decrease_values = torch.flip(decrease_values, dims=(0,))
# below given state
for i in range(num_steps_down):
updated_tensor = og_tensor.clone()
updated_tensor[0, 0] -= decrease_values[i]
updated_tensor[0, 1:] += (decrease_values[i] / (og_tensor.size(1) - 1))
if torch.any(torch.lt(updated_tensor, 0)):
break
elif torch.equal(updated_tensor.squeeze(), torch.Tensor(w_optimal)):
continue
else:
result_list.append(updated_tensor.squeeze())
return result_list
In [62]:
# og_tensor = torch.tensor([[0.1, 0.3, 0.3, 0.3]])
high_dim_tensors = get_high_dim_tensors(og_tensor=optimal_weight, w_optimal=optimal_weight, num_steps=20)
high_dim_tensors[:5]
Out[62]:
[tensor([0.3265, 0.2698, 0.2198, 0.1839]), tensor([1., 0., 0., 0.]), tensor([0.0000, 0.3333, 0.3333, 0.3333]), tensor([0.2500, 0.2500, 0.2500, 0.2500]), tensor([0.4013, 0.2449, 0.1949, 0.1589])]
In [63]:
delta_weights_0 = [0] + [(-high_dim_tensors[0][0] + high_dim_tensors[1:][i][0]).item() for i in range(len(high_dim_tensors) - 1)]
delta_weights_0 = [round(i, 2) for i in delta_weights_0]
delta_weights_0[:5]
Out[63]:
[0, 0.67, -0.33, -0.08, 0.07]
In [64]:
tmp = high_dim_tensors
tmp = [i.detach().numpy() for i in tmp]
df = pd.DataFrame(tmp).T
px.line(df).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [65]:
# out = []
# for i in high_dim_tensors:
# qval = a(torch.FloatTensor(i)).detach().numpy()
# act = env.action_space[qval.argmax()]
# out.append(act.item())
# if act.item() == 1:
# print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
# else:
# print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
In [66]:
out = {}
for tc in tc_list:
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
tmp = []
for i in high_dim_tensors:
qval = a(torch.FloatTensor(i)).detach().numpy()
act = env.action_space[qval.argmax()]
tmp.append(act.item())
out[f'TC: {tc * 1e4:.0f} bps'] = tmp
... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ...
In [67]:
df = pd.DataFrame(out, index=delta_weights_0)
df = df.sort_index()
df.head()
Out[67]:
| TC: 0 bps | TC: 5 bps | TC: 10 bps | TC: 20 bps | TC: 50 bps | TC: 100 bps | TC: 500 bps | |
|---|---|---|---|---|---|---|---|
| -0.33 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| -0.30 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| -0.22 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| -0.15 | 1 | 1 | 1 | 0 | 0 | 0 | 0 |
| -0.08 | 1 | 1 | 1 | 0 | 0 | 0 | 0 |
In [68]:
sns.heatmap(df, cmap='cividis')
ax = plt.gca()
plt.hlines(y=df.shape[0] // 2 - 1/2, xmin=0, xmax=len(dqn_action_df.columns), colors='red')
plt.title("DQN Binary Heatmap of Rebalancing")
plt.ylabel("Asset 1 Delta")
plt.show()
In [69]:
# Assuming `dqn_bool_action_df` and `bool_action_bm_df` are your DataFrames
dqn_df = df.copy()
# Create a heatmap trace for DQN Binary
trace_dqn = go.Heatmap(z=dqn_df, x=dqn_df.columns, y=dqn_df.index, name="DQN Binary")
# Create the layout with 'overlay' parameter set to "overlay"
layout = go.Layout(
title=f"DQN Binary for {n_asset} case",
xaxis=dict(title="Transaction Cost (bps)",
showgrid=True),
yaxis=dict(title="Delta Weight on First Asset",
showgrid=True))
# Add horizontal line at y=0.50
layout.shapes = [
dict(
type="line",
x0=dqn_df.columns[0], # Adjust x0 to the starting point of your data
x1=dqn_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
)
]
fig = go.Figure(data=[trace_dqn], layout=layout)
fig.show()
In [70]:
dqn_noreb = {}
for col in df.columns:
dqn_index = df[df[col] == 0].index.values
dqn_noreb[col] = dqn_index
In [71]:
# Create a DataFrame from the dictionary, excluding keys with empty values
df = pd.DataFrame.from_dict(dqn_noreb, orient='index').T
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Delta Weight Asset 1")
# Show the plot
fig.show()
In [72]:
df = optimal_weight[0] + df
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Weight Asset 1")
# Show the plot
fig.show()
8 Assets¶
In [73]:
n_asset = 8
tc = 0.002
mu = np.linspace(50, 200, n_asset) / 1e4
sigma = np.linspace(300, 800, n_asset) / 1e4
cov = np.diag(sigma ** 2)
In [74]:
optimal_weight = find_optimal_wgt(mu, cov)
state = torch.Tensor(optimal_weight).unsqueeze(0)
optimal_weight
Out[74]:
array([0.16397001, 0.15284272, 0.13978576, 0.12758588, 0.11680825,
0.10743797, 0.09931577, 0.09225364])
In [75]:
px.line(optimal_weight).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [76]:
tc = 0.002
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
... loading checkpoint ... ... loading checkpoint ...
In [77]:
qval = a(torch.FloatTensor(state)).detach().numpy()
act = env.action_space[qval.argmax()]
if act.item() == 1:
print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
else:
print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
Action = 0 -> You should NOT rebalance! (tc = 20 bps)
In [78]:
# og_tensor = torch.tensor([[0.1, 0.3, 0.3, 0.3]])
high_dim_tensors = get_high_dim_tensors(og_tensor=optimal_weight, w_optimal=optimal_weight, num_steps=20)
high_dim_tensors[:5]
Out[78]:
[tensor([0.1640, 0.1528, 0.1398, 0.1276, 0.1168, 0.1074, 0.0993, 0.0923]), tensor([1., 0., 0., 0., 0., 0., 0., 0.]), tensor([0.0000, 0.1429, 0.1429, 0.1429, 0.1429, 0.1429, 0.1429, 0.1429]), tensor([0.1250, 0.1250, 0.1250, 0.1250, 0.1250, 0.1250, 0.1250, 0.1250]), tensor([0.2569, 0.1396, 0.1265, 0.1143, 0.1035, 0.0942, 0.0860, 0.0790])]
In [79]:
delta_weights_0 = [0] + [(-high_dim_tensors[0][0] + high_dim_tensors[1:][i][0]).item() for i in range(len(high_dim_tensors) - 1)]
delta_weights_0 = [round(i, 2) for i in delta_weights_0]
delta_weights_0[:5]
Out[79]:
[0, 0.84, -0.16, -0.04, 0.09]
In [80]:
tmp = high_dim_tensors
tmp = [i.detach().numpy() for i in tmp]
df = pd.DataFrame(tmp).T
px.line(df).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [81]:
# out = []
# for i in high_dim_tensors:
# qval = a(torch.FloatTensor(i)).detach().numpy()
# act = env.action_space[qval.argmax()]
# out.append(act.item())
# if act.item() == 1:
# print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
# else:
# print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
In [82]:
out = {}
for tc in tc_list:
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
tmp = []
for i in high_dim_tensors:
qval = a(torch.FloatTensor(i)).detach().numpy()
act = env.action_space[qval.argmax()]
tmp.append(act.item())
out[f'TC: {tc * 1e4:.0f} bps'] = tmp
... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ...
In [83]:
df = pd.DataFrame(out, index=delta_weights_0)
df = df.sort_index()
df.head()
Out[83]:
| TC: 0 bps | TC: 5 bps | TC: 10 bps | TC: 20 bps | TC: 50 bps | TC: 100 bps | TC: 500 bps | |
|---|---|---|---|---|---|---|---|
| -0.16 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
| -0.09 | 1 | 1 | 1 | 0 | 0 | 0 | 0 |
| -0.04 | 1 | 0 | 1 | 1 | 0 | 0 | 0 |
| 0.00 | 1 | 1 | 1 | 0 | 0 | 0 | 0 |
| 0.09 | 1 | 1 | 1 | 1 | 0 | 0 | 0 |
In [84]:
sns.heatmap(df, cmap='cividis')
ax = plt.gca()
plt.hlines(y=df.shape[0] // 2 - 1/2, xmin=0, xmax=len(dqn_action_df.columns), colors='red')
plt.title("DQN Binary Heatmap of Rebalancing")
plt.ylabel("Asset 1 Delta")
plt.show()
In [85]:
# Assuming `dqn_bool_action_df` and `bool_action_bm_df` are your DataFrames
dqn_df = df.copy()
# Create a heatmap trace for DQN Binary
trace_dqn = go.Heatmap(z=dqn_df, x=dqn_df.columns, y=dqn_df.index, name="DQN Binary")
# Create the layout with 'overlay' parameter set to "overlay"
layout = go.Layout(
title=f"DQN Binary for {n_asset} case",
xaxis=dict(title="Transaction Cost (bps)",
showgrid=True),
yaxis=dict(title="Delta Weight on First Asset",
showgrid=True))
# Add horizontal line at y=0.50
layout.shapes = [
dict(
type="line",
x0=dqn_df.columns[0], # Adjust x0 to the starting point of your data
x1=dqn_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
)
]
fig = go.Figure(data=[trace_dqn], layout=layout)
fig.show()
In [86]:
dqn_noreb = {}
for col in df.columns:
dqn_index = df[df[col] == 0].index.values
dqn_noreb[col] = dqn_index
In [87]:
# Create a DataFrame from the dictionary, excluding keys with empty values
df = pd.DataFrame.from_dict(dqn_noreb, orient='index').T
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Delta Weight Asset 1")
# Show the plot
fig.show()
In [88]:
df = optimal_weight[0] + df
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Weight Asset 1")
# Show the plot
fig.show()
32 Assets¶
In [89]:
n_asset = 32
tc = 0.002
mu = np.linspace(50, 200, n_asset) / 1e4
sigma = np.linspace(300, 800, n_asset) / 1e4
cov = np.diag(sigma ** 2)
In [90]:
optimal_weight = find_optimal_wgt(mu, cov)
state = torch.Tensor(optimal_weight).unsqueeze(0)
optimal_weight
Out[90]:
array([0.04102786, 0.04066222, 0.04002113, 0.03934691, 0.038726 ,
0.03804007, 0.03724126, 0.03634512, 0.03555705, 0.03485426,
0.03422818, 0.03350525, 0.03282634, 0.0321328 , 0.03141941,
0.03083873, 0.03032973, 0.02962892, 0.02909339, 0.02865538,
0.02816096, 0.02748528, 0.02704373, 0.0265696 , 0.02598976,
0.02559782, 0.02520074, 0.02475413, 0.0242343 , 0.02387018,
0.02346884, 0.02314463])
In [91]:
px.line(optimal_weight).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [92]:
tc = 0.002
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
... loading checkpoint ... ... loading checkpoint ...
In [93]:
qval = a(torch.FloatTensor(state)).detach().numpy()
act = env.action_space[qval.argmax()]
if act.item() == 1:
print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
else:
print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
Action = 1 -> You should rebalance! (tc = 20 bps)
In [94]:
# og_tensor = torch.tensor([[0.1, 0.3, 0.3, 0.3]])
high_dim_tensors = get_high_dim_tensors(og_tensor=optimal_weight, w_optimal=optimal_weight, num_steps=20)
# high_dim_tensors[:5]
In [95]:
delta_weights_0 = [0] + [(-high_dim_tensors[0][0] + high_dim_tensors[1:][i][0]).item() for i in range(len(high_dim_tensors) - 1)]
delta_weights_0 = [round(i, 2) for i in delta_weights_0]
delta_weights_0[:5]
Out[95]:
[0, 0.96, -0.04, -0.01, 0.11]
In [96]:
tmp = high_dim_tensors
tmp = [i.detach().numpy() for i in tmp]
df = pd.DataFrame(tmp).T
px.line(df).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [97]:
# out = []
# for i in high_dim_tensors:
# qval = a(torch.FloatTensor(i)).detach().numpy()
# act = env.action_space[qval.argmax()]
# out.append(act.item())
# if act.item() == 1:
# print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
# else:
# print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
In [98]:
out = {}
for tc in tc_list:
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
tmp = []
for i in high_dim_tensors:
qval = a(torch.FloatTensor(i)).detach().numpy()
act = env.action_space[qval.argmax()]
tmp.append(act.item())
out[f'TC: {tc * 1e4:.0f} bps'] = tmp
... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ...
In [99]:
df = pd.DataFrame(out, index=delta_weights_0)
df = df.sort_index()
df.head()
Out[99]:
| TC: 0 bps | TC: 5 bps | TC: 10 bps | TC: 20 bps | TC: 50 bps | TC: 100 bps | TC: 500 bps | |
|---|---|---|---|---|---|---|---|
| -0.04 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| -0.01 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.00 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.11 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.21 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
In [100]:
sns.heatmap(df, cmap='cividis')
ax = plt.gca()
plt.hlines(y=df.shape[0] // 2 - 1/2, xmin=0, xmax=len(dqn_action_df.columns), colors='red')
plt.title("DQN Binary Heatmap of Rebalancing")
plt.ylabel("Asset 1 Delta")
plt.show()
In [101]:
# Assuming `dqn_bool_action_df` and `bool_action_bm_df` are your DataFrames
dqn_df = df.copy()
# Create a heatmap trace for DQN Binary
trace_dqn = go.Heatmap(z=dqn_df, x=dqn_df.columns, y=dqn_df.index, name="DQN Binary")
# Create the layout with 'overlay' parameter set to "overlay"
layout = go.Layout(
title=f"DQN Binary for {n_asset} case",
xaxis=dict(title="Transaction Cost (bps)",
showgrid=True),
yaxis=dict(title="Delta Weight on First Asset",
showgrid=True))
# Add horizontal line at y=0.50
layout.shapes = [
dict(
type="line",
x0=dqn_df.columns[0], # Adjust x0 to the starting point of your data
x1=dqn_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
)
]
fig = go.Figure(data=[trace_dqn], layout=layout)
fig.show()
In [102]:
dqn_noreb = {}
for col in df.columns:
dqn_index = df[df[col] == 0].index.values
dqn_noreb[col] = dqn_index
In [103]:
# Create a DataFrame from the dictionary, excluding keys with empty values
df = pd.DataFrame.from_dict(dqn_noreb, orient='index').T
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Delta Weight Asset 1")
# Show the plot
fig.show()
In [104]:
df = optimal_weight[0] + df
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Weight Asset 1")
# Show the plot
fig.show()
128 Assets¶
In [105]:
n_asset = 128
tc = 0.002
mu = np.linspace(50, 200, n_asset) / 1e4
sigma = np.linspace(300, 800, n_asset) / 1e4
cov = np.diag(sigma ** 2)
In [106]:
optimal_weight = find_optimal_wgt(mu, cov)
state = torch.Tensor(optimal_weight).unsqueeze(0)
optimal_weight
Out[106]:
array([0.01030035, 0.01026876, 0.01023511, 0.0102 , 0.01015838,
0.01011766, 0.01008988, 0.01006437, 0.01003273, 0.00998191,
0.00993001, 0.00986722, 0.0098301 , 0.009819 , 0.00980393,
0.00966758, 0.00971091, 0.00966272, 0.0096263 , 0.00957051,
0.00951943, 0.00947479, 0.00943105, 0.00938339, 0.00933417,
0.00927284, 0.00923469, 0.00919609, 0.00915945, 0.00911217,
0.00906382, 0.00901939, 0.00898043, 0.00892036, 0.00887741,
0.00883667, 0.00875916, 0.00873536, 0.00870878, 0.00866395,
0.00858791, 0.0085759 , 0.00851901, 0.00849197, 0.00845346,
0.00838184, 0.00836867, 0.00831132, 0.00825751, 0.00824467,
0.00820846, 0.00812636, 0.00810293, 0.00805285, 0.00804493,
0.0079946 , 0.00795561, 0.00792474, 0.0078701 , 0.00784109,
0.0077971 , 0.00775392, 0.00770886, 0.00768773, 0.00762451,
0.0075941 , 0.00755294, 0.00749941, 0.00747001, 0.00743369,
0.00739891, 0.00737036, 0.0073371 , 0.00730792, 0.00726789,
0.00723004, 0.00720295, 0.0071536 , 0.00713352, 0.00711362,
0.00705915, 0.00702382, 0.00699875, 0.00696341, 0.00694701,
0.00690769, 0.00686863, 0.00683495, 0.00680933, 0.00676929,
0.00675661, 0.00671444, 0.00669606, 0.00666794, 0.00664377,
0.00658449, 0.00657847, 0.00652719, 0.00651279, 0.00649585,
0.0064654 , 0.00643028, 0.00638831, 0.00638449, 0.00635961,
0.00632476, 0.00628446, 0.00627887, 0.00623705, 0.00620711,
0.00618739, 0.00616401, 0.00614963, 0.00610359, 0.00609171,
0.00608517, 0.00604394, 0.00602379, 0.00599833, 0.00597261,
0.00594517, 0.00593802, 0.00588976, 0.0058723 , 0.00585486,
0.00581072, 0.00579242, 0.00578363])
In [107]:
px.line(optimal_weight).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [108]:
tc = 0.002
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
... loading checkpoint ... ... loading checkpoint ...
In [109]:
qval = a(torch.FloatTensor(state)).detach().numpy()
act = env.action_space[qval.argmax()]
if act.item() == 1:
print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
else:
print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
Action = 1 -> You should rebalance! (tc = 20 bps)
In [110]:
# og_tensor = torch.tensor([[0.1, 0.3, 0.3, 0.3]])
high_dim_tensors = get_high_dim_tensors(og_tensor=optimal_weight, w_optimal=optimal_weight, num_steps=20)
# high_dim_tensors[:5]
In [111]:
delta_weights_0 = [0] + [(-high_dim_tensors[0][0] + high_dim_tensors[1:][i][0]).item() for i in range(len(high_dim_tensors) - 1)]
delta_weights_0 = [round(i, 2) for i in delta_weights_0]
delta_weights_0[:5]
Out[111]:
[0, 0.99, -0.01, -0.0, 0.11]
In [112]:
tmp = high_dim_tensors
tmp = [i.detach().numpy() for i in tmp]
df = pd.DataFrame(tmp).T
px.line(df).update_layout(xaxis_title="Asset i", yaxis_title="Weight in Asset i")
In [113]:
# out = []
# for i in high_dim_tensors:
# qval = a(torch.FloatTensor(i)).detach().numpy()
# act = env.action_space[qval.argmax()]
# out.append(act.item())
# if act.item() == 1:
# print(f"Action = {act.item()} -> You should rebalance! (tc = {tc * 1e4:.0f} bps)")
# else:
# print(f"Action = {act.item()} -> You should NOT rebalance! (tc = {tc * 1e4:.0f} bps)")
In [114]:
out = {}
for tc in tc_list:
a = DQN(n_observations=n_asset, n_actions=2, name="Policy", tc=tc).to(tmp_device)
b = DQN(n_observations=n_asset, n_actions=2, name="Target", tc=tc).to(tmp_device)
load_models(pn=a, tn=b)
tmp = []
for i in high_dim_tensors:
qval = a(torch.FloatTensor(i)).detach().numpy()
act = env.action_space[qval.argmax()]
tmp.append(act.item())
out[f'TC: {tc * 1e4:.0f} bps'] = tmp
... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ... ... loading checkpoint ...
In [120]:
df = pd.DataFrame(out, index=delta_weights_0)
df = df.sort_index()
df.head()
Out[120]:
| TC: 0 bps | TC: 5 bps | TC: 10 bps | TC: 20 bps | TC: 50 bps | TC: 100 bps | TC: 500 bps | |
|---|---|---|---|---|---|---|---|
| -0.01 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.00 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| -0.00 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
| 0.11 | 1 | 1 | 1 | 1 | 1 | 1 | 0 |
| 0.22 | 1 | 1 | 1 | 1 | 1 | 1 | 0 |
In [121]:
sns.heatmap(df, cmap='cividis')
ax = plt.gca()
plt.hlines(y=df.shape[0] // 2 - 1/2, xmin=0, xmax=len(dqn_action_df.columns), colors='red')
plt.title("DQN Binary Heatmap of Rebalancing")
plt.ylabel("Asset 1 Delta")
plt.show()
In [122]:
# Assuming `dqn_bool_action_df` and `bool_action_bm_df` are your DataFrames
dqn_df = df.copy()
# Create a heatmap trace for DQN Binary
trace_dqn = go.Heatmap(z=dqn_df, x=dqn_df.columns, y=dqn_df.index, name="DQN Binary")
# Create the layout with 'overlay' parameter set to "overlay"
layout = go.Layout(
title=f"DQN Binary for {n_asset} case",
xaxis=dict(title="Transaction Cost (bps)",
showgrid=True),
yaxis=dict(title="Delta Weight on First Asset",
showgrid=True))
# Add horizontal line at y=0.50
layout.shapes = [
dict(
type="line",
x0=dqn_df.columns[0], # Adjust x0 to the starting point of your data
x1=dqn_df.columns[-1], # Adjust x1 to the ending point of your data
y0=optimal_weight[0],
y1=optimal_weight[0],
line=dict(color="red", width=2), # You can customize the color and width of the line
)
]
fig = go.Figure(data=[trace_dqn], layout=layout)
fig.show()
In [117]:
dqn_noreb = {}
for col in df.columns:
dqn_index = df[df[col] == 0].index.values
dqn_noreb[col] = dqn_index
In [118]:
# Create a DataFrame from the dictionary, excluding keys with empty values
df = pd.DataFrame.from_dict(dqn_noreb, orient='index').T
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Delta Weight Asset 1")
# Show the plot
fig.show()
In [119]:
df = optimal_weight[0] + df
# Create a box plot using Plotly
fig = px.box(df, title="No Rebalance Range").update_layout(yaxis_title="Weight Asset 1")
# Show the plot
fig.show()
In [ ]:
In [ ]:
In [ ]: